Scan-to-QC: Quality Control of Prefabricated Structural Systems Using Point Clouds

Background: The use of point clouds to perform quality control (QC) of prefabricated structural systems is growing markedly. The sustained application of point clouds on construction projects at scale, however, requires reliable and automated frameworks to foster practicality. Literature Review: Existing solutions seldom incorporate the impact of instrument artifacts, construction errors, and complex element geometries, rendering the frameworks error-prone and sensitive to thresholds, outliers, and missing data. More specifically, current methods: 1) are not generalizable to any freeform element; 2) rely on local primitives, which increases runtime and threshold subjectivity; 3) do not formulate the point cloud to model fitting based on geometric distance; and 4) do not incorporate a dedicated outlier detection to explicitly handle large outliers. Methodology: This study presents, scan-to-QC, a generic solution for automatic QC of prefabricated elements with non-analytic cross-sections using point clouds. The framework addresses the current state-of-the-art limitations (above) by providing a generic, direct artificial-intelligence (AI)-based geometric distance fitting robust to outliers. ... mehrThe proposed approach introduces four new algorithms to: 1) represent the cross-sectional boundaries of elements with trigonometric polynomials, 2) perform orthogonal fitting of points to these polynomials, 3) remove outlying observations through robust concentration steps (CS), and 4) perform assembly verification of prefabricated elements in construction projects. As the convergence of CS depends on the initial fitting parameters, the performances of seven deterministic, heuristic, and AI-based (i.e., metaheuristic and global computational intelligence-based) optimization strategies to recover these parameters were evaluated on point clouds of five common prefabricated elements with complex geometries. The comparative analysis considered factors such as the fit quality, object detection accuracy, and computation runtime. Results and Conclusions: Among these methods, the Particle Swarm (PS) achieved the best results, significantly outperforming Simulated Annealing, Bayesian, and Genetic Algorithm. Furthermore, by fine-tuning the swarm size, the PS attained 99% and 96% accuracies in object detection and fitting quality, respectively. Finally, the calibrated PS algorithm was applied to the point cloud of a real-world prefabricated spaceframe structure, yielding 98.5% reliability in nodal coordinate estimation compared to the ground truth.